retail-data-v0.1.0.tar.gz dict | retail-sft-v0.1.0.tar.gz dict | verl-source-v0.1.0.tar.gz dict |
|---|---|---|
{
"bytes": 23611843,
"sha256": "0997a01a6d9b49df65998d565b8a6ae175d06a99730c6a45e356ab8fa4db9235",
"files": 46,
"parts": [
{
"filename": "bundles/retail-data-v0.1.0.000.bin",
"bytes": 2097152,
"sha256": "26e6f75a7ca33dc4fa2be2c1488e76f4b83aba400af32f404c6c6924307bdd1e"
},
{
"... | {
"bytes": 47856904,
"sha256": "655822bde3cf38ea3f39c884691f5216f49e705f604bc1a54a68a448acbb01ec",
"files": 13,
"parts": [
{
"filename": "bundles/retail-sft-v0.1.0.000.bin",
"bytes": 2097152,
"sha256": "5ed9e4f399534fcf040dd2f12ec1c38f3bd31e9c534971d44608e245c31a7ac0"
},
{
"f... | {
"bytes": 2945110,
"sha256": "9357ded098bb9485548d3316ac79864448d64a943208a02afd9bdeca66123963",
"files": 1252,
"parts": [
{
"filename": "bundles/verl-source-v0.1.0.000.bin",
"bytes": 2097152,
"sha256": "12cb775e662ddc63dce1f247467192e72158227b4996cc75d95cd6ef5fc4d0be"
},
{
... |
RetailAgent-RL artifacts — v0.1.0
Frozen data, original SFT adapter and verl source snapshot for lbw-work/retail-agent-rl.
| File | Contents |
|---|---|
| retail-data-v0.1.0.tar.gz | retail_full_v2 SFT/RL splits, retail_rl_v3 94/16/22 frozen task view and quarantine reports, four synthetic Retail databases, tokenizer and provenance |
| retail-sft-v0.1.0.tar.gz | Existing Qwen3-4B SFT LoRA adapter, tokenizer, run metadata and evaluation manifest; no optimizer pickle files |
| verl-source-v0.1.0.tar.gz | Frozen local verl source used by the experimental migration; source fingerprint is in GitHub configs/verl/source_lock.json |
Transport uses 2 MiB binary chunks under bundles/; bundle_checksums.json records every part and the complete archive SHA256. The downloader reassembles and verifies the archive before extraction. Bundles retain paths relative to the Retail repository. Use scripts/fetch_artifacts.py from GitHub to download, verify SHA256 and restore them. Existing conflicting files are rejected.
Provenance and scope
Data derives from inclusionAI/AReaL-tau2-data, revision 86971dc03da6e7c1a7933295e05b84aab8215386 (Apache-2.0). Base model: Qwen/Qwen3-4B, revision 1cfa9a7208912126459214e8b04321603b3df60c (Apache-2.0). The original SFT adapter was trained on 5,682 train examples; it is distinct from the community GRPO checkpoint.
The environment is sierra-research/tau2-bench, commit f8de30c298689cbe0117d76a378e7315a17e5bd8 (MIT). Bundled verl source is Apache-2.0; preserve its LICENSE and Notice.txt. Each upstream component retains its own license. These are synthetic benchmark scenarios, not real customer records.
Important limitations
This is a development snapshot. The frozen task pool is not a semantic correctness guarantee. retail_135 has unresolved task/color alignment and an upstream multi-item update defect. retail_136 has a customer-confirmed payment choice that differs from the single gold state. Original rewards are retained; these tasks require review before further training. No full-pool GRPO improvement or held-out community-model result is claimed. Source data and prepared Arrow caches are not duplicated; the upstream source link and fixed revision are retained.
- Downloads last month
- 88